Chairs: | Paul Gustafson, Victor Kipnis, Pamela Shaw |
Members: | Jonathan Bartlett, Hendriek Boshuizen, Raymond Carroll, Veronika Deffner, Kevin Dodd, Laurence Freedman, Sabine Hoffmann, Ruth Keogh, Helmut Küchenhoff, Douglas Midthune, Cécile Proust-Lima, Anne Thiebaut, Janet Tooze, Michael Wallace |
Homepage: Topic Group 4
Measurement error and misclassification, hereafter MEM, in covariates and responses occurs in many observational studies and some experimental studies. See Buonaccorsi 2010^{1}, Carroll 2006^{2} and Gustafson 2004^{3} for textbook treatments of the literature.
MEM can be taken as an extreme version of a missing data problem, where the true data affected by error is missing in all or the vast majority of subjects. This special feature requires special methodology and means of thinking.
It is well-known that MEM in predictors can result in biased parameter estimates and a loss of statistical power, sometimes a startling loss of power. Not accounting for MEM can lead to over-optimistic power calculations and result in failed studies. It is less well-known that measurement error can, in some circumstances, lead to incorrect inferences and hypothesis testing, i.e., it is not always the case that MEM results in a bias towards the null.
We will provide guidance into the following series of topics. The listing is not in order of importance.
When will MEM likely affect the validity of statistical analysis?
When will MEM result simply in a bias towards the null, so that parameter estimates are biased by hypothesis testing for null effects is valid but less powerful than if there were no MEM? Conversely, when will MEM result in bias that is not towards the null, so that incorrect inferences and conclusions will result?
How can one design a study in the presence of MEM both to assess their extent and to provide properly tuned sample sizes for sufficient statistical power?
What software is available to make sample size calculations when there is measurement error or misclassification in the key predictor?
What is the role of Bayesian analysis in MEM?
What software is available for a measurement error analysis? We will give advice on available software for measurement error analysis. Here is a sampling.
In SAS, the
CALIS
procedure, SAS macros available from the National Cancer Institute programs from Iowa State University, and programs from the Harvard School of Public Health, are available. In R, there are the decon and simex packages. Stata also has programs for regression calibration.What analyses are possible if some variables are subject to measurement error and others are subject to misclassification?
What can be done if measurement error or misclassification is known to exist but data are not available in a current study to assess the impacts of these errors?
How can one perform variable selection when some predictors are subject to MEM?
What can be done if there is MEM in the response variable?
Buonaccorsi JP. Measurement Error: Models, Methods, and Applications. Chapman & Hall/CRC: Boca Raton, Florida, 2010.
Carroll RJ, Ruppert D, Stefanski LA, Crainiceanu CM. Measurement Error in Nonlinear Models: A Modern Perspective, Second Edition. Chapman and Hall/CRC Press: Boca Raton, Florida, 2006.
Gustafson P. Measurement Error and Misclassification in Statistics and Epidemiology. Chapman&Hall/CRC: Boca Raton, Florida, 2004.